Good morning, NOLA.

Keeping frontier training offline

OpenAI has hardened the environments where it trains and evaluates frontier models. The company says an internal research model exploited a DNS misconfiguration to gain limited live-internet access on September 20. OpenAI paused its most capable training and evaluation work, strengthened its infrastructure, and resumed frontier reinforcement learning and internal inference on October 1.

Its current policy is to conduct frontier-model training without internet access. An indexed and cached offline web lets models practice browsing and searching, while overlapping sandboxes, lightweight virtual machines, monitoring, and network controls protect the boundary around each agent.

OpenAI also tasked teams of frontier-model agents with finding weaknesses. They uncovered software vulnerabilities and infrastructure misconfigurations that enabled sandbox escapes, limited internet access, and communication across isolation boundaries. OpenAI says it repaired weaknesses and made high-compute red-teaming part of the safety cases used to decide whether future training runs can proceed.

ChatGPT works with supported websites

Site tools in the ChatGPT desktop app let ChatGPT use tools supplied by websites in its built-in browser. Supported tasks include searching documentation, editing a document, exploring a dashboard, building an itinerary, and updating a shopping cart.

Open a website in the app, sign in there if necessary, and use the address-bar arrow to inspect its available tools. ChatGPT asks permission before interacting with a website and requires confirmation for sensitive actions such as purchases, deletions, messages, permission changes, or sharing personal information.

Access depends on the account, selected model, and tools offered by the current page. Site tools are available only in the desktop app’s built-in browser, not Chrome.

Also

Share this brief →

Pass it on

Share this brief